MBA candidate, USC Marshall · Former ML Team Lead
Ankit Kapoor
I build working products for hard business questions, and the strategy behind them.
- person gen-AI team built from scratch
- 15
- faster medical coding per chart
- 10×
- CPT test accuracy on surgical charts
- 91%
At Ascent Business Solutions, building AI for U.S. medical coding.
01 — Work
15 products, ready to explore.
Explore browser demos and downloadable apps. Open a project, read the code, or go deeper in a case study.
Market Entry War Room
Ranks 393 US metro areas for a business's next location, then tests whether the winner survives a change in priorities.
Why it mattersMarket-entry advice usually arrives as one answer. This shows which assumption is actually driving the pick.
- Python
- FastAPI
- React
- Census ACS
- Monte Carlo
The Exception Engine
A policy lab built on EU air-passenger rights. Code applies the rules, an AI model answers bounded factual questions, and a confidence threshold decides when a person steps in.
Why it mattersMost automation decisions live in the grey area. This makes the trade-off between automating and asking a human explicit, case by case, across 600 recorded model assessments.
- TypeScript
- Vite
- Cloudflare Workers
- TypeSafe JEV
- EU 261/2004
ValueLens
Rebuilds a company's cash flows from its SEC filings and reads the share price three ways: market-implied, consensus, and fundamental.
Why it mattersTurns “is this stock cheap?” into “what growth is the price betting on?”, with value ranges calibrated on held-out data.
- JavaScript
- Cloudflare Workers
- SEC EDGAR / XBRL
- DCF
- Monte Carlo
The Escalation Frontier
Routes 1,270 recorded banking-support decisions between a fast classifier, a reasoning model, and human review. Move the confidence threshold and watch cost, accuracy, and latency shift.
Why it mattersNot every decision deserves the same model. This puts a price on extra certainty, so the routing threshold becomes a business choice instead of a guess.
- TypeScript
- React
- Node.js
- TypeSafe JEV
- Claude Sonnet 4.5
- BANKING77
OrthoCode AI
Medical coding for orthopedic surgery. Every proposed code links back to the chart page that justifies it, and CMS rules are checked in code.
Why it mattersAt Ascent, my team cut coding time per chart 10× by targeting surgical charts instead of radiology. OrthoCode rebuilds that bet so every code can survive an audit.
- Python
- FastAPI
- Next.js
- PostgreSQL
- OpenAI Structured Outputs
tandem
A Mac app that puts you, Claude Code, and Codex in one shared conversation, with Solo, Relay, Parallel, and Debate modes.
Why it mattersA second perspective is useful only if it has the context. tandem lets the agents read each other’s replies, challenge assumptions, and work in an order you control.
- Electron
- Node.js
- Claude Code
- Codex CLI
Narrative vs. Numbers
Reads the priorities management states in a 10-K, then checks whether the spending in its filings backs them up.
Why it mattersEvery strategy deck says the right things. The gap between what a company says and what it does is where execution shows.
- Python
- SEC XBRL
- OpenAI
- Deterministic scoring
Catalog Underwriter
Values a music catalog from public listening data, with an editable DCF, bear/base/bull cases, and a “break the deal” stress test.
Why it mattersCatalog deals are priced on decay assumptions. This shows exactly which ones break the return.
- React
- TypeScript
- Recharts
- MusicBrainz
- Last.fm
- YouTube API
Bill Buster
Turns a hospital bill into a cited inquiry letter and a phone checklist. Everything runs in the browser, so the bill never leaves the device.
Why it mattersHospital prices are now published, but patients still can't use them. This turns the data into specific questions to ask.
- JavaScript
- PDF.js
- pdf-lib
- CMS NCCI & MUE
- Cloudflare Workers
The Moat Test
Builds a cheap challenger to a paid AI product (meeting assistants) and measures what the subscription is really buying.
Why it mattersIf a model can do the feature, the moat has to be somewhere else. This is an experiment to find where.
- Next.js
- TypeScript
- LLM extraction
- Baseline benchmarks
Priced In
Runs a valuation backwards: what growth, margin, and customer count would justify the price, and what an AI bet must contribute.
Why it mattersReframes an AI investment from “will it pay off?” to “what must it deliver for the valuation to hold?”
- Next.js
- TypeScript
- SEC filings
- decimal.js
Disrupt This Business
A four-quarter strategy game: defend a software incumbent against an AI-native challenger, then switch sides and attack your own plan.
Why it mattersIncumbents rarely lose for lack of technology. They lose defending the wrong part of the business, and this makes that trade-off playable.
- Next.js
- TypeScript
- Economic engine
- Playwright
FlyCast
Uses the fruit-fly connectome (166,691 real neurons) as a fixed recurrent network and tests whether it can forecast time series.
Why it mattersAsks whether evolved wiring computes better than random wiring, using shuffled-weight controls and no data leakage.
- Python
- NumPy
- SciPy sparse
- scikit-learn
- Next.js
Ideology Inventory
A political-orientation survey with separate economic, social, and authority scales, moral-foundation profiles, and response-quality grading.
Why it mattersMost online quizzes squash views onto one axis. This measures them the way survey researchers do, benchmarked against ANES and CES.
- TypeScript
- Vite
- Cloudflare D1
- R (IRT)
Wordle Solver
Type the word you played and the colors Wordle gave back. It narrows 5,757 possible answers and suggests the best next guesses.
Why it mattersA small one, built for fun: elimination plus letter-frequency scoring, ported from Python to run entirely in the browser.
- JavaScript
- Python
- Cloudflare Workers
02 — About
Engineer first. Strategist by training.
Experience
Mar 2023 — Jul 2026
Machine Learning Engineer → Team Lead
Ascent Business Solutions · Nagpur, India
Medical coding and revenue-cycle services for U.S. providers
- Built and managed a 15-person generative AI team from scratch, partnering with HR on hiring, to deliver a portfolio of concurrent healthcare revenue-cycle products.
- Cut coding time per chart 10× at up to 91% CPT test accuracy by focusing the auto-coding product on orthopedic and urology surgery charts over easier radiology charts, for their larger implementation scope.
- Shipped a Clinical Documentation Improvement product for U.S. hospitals that checked chart completeness against payer documentation requirements and flagged errors likely to trigger claim denials.
- Turned workflow bottlenecks into product requirements with medical-coding, operations, and client teams.
Oct 2022 — Mar 2023
Conversational AI Engineer Trainee
Bradsol · Hyderabad, India
- Led a four-person team that turned client requirements into an HR conversational-AI chatbot built on Kore.ai.
- Proposed generative-AI synthetic training data, which was adopted to improve the chatbot's free-form conversation.
Education
Expected May 2028
Master of Business Administration
University of Southern California, Marshall School of Business
High-Tech Association · Competitive Strategy · Organizational Behavior and Leadership
2022
Bachelor of Engineering (Hons.)
Birla Institute of Technology & Science, Pilani — Dubai
Computer Science
Publication
“Implementing Natural Language Generation Through Industry-Specific Chatbots”
Ankit Kapoor, Sujala Shetty · Soft Computing Research Society (SCRS), 2023
Built a GPT-2 medical-query chatbot trained on about 45,000 conversational exchanges to show how text generation fixes the rigidity of template chatbots.
Read the paper (PDF) (opens in a new tab)Skills
Product & strategy
- AI product strategy
- Competitive strategy
- Market-entry analysis
- Financial modeling
- Experimentation
Technical & AI
- Python
- SQL
- PostgreSQL
- FastAPI
- Docker
- OpenAI API
- AWS Bedrock
- NLP
Agentic tools
- Claude Code
- Codex
- Cursor
- OpenClaw
03 — Contact
Let's build something.
Open to conversations about strategy, product, and AI roles, and happy to walk through any project above.
Thanks for scrolling this far —